The Moon's surface is heavily cratered, these features serving as invaluable records of the Solar System's history. Existing lunar crater databases, compiled through visual inspection and automated detection methods, suffer from inconsistencies in crater counts due to subjective manual analysis and limitations in the automated methods. The International Astronomical Union (IAU) has recognized 9137 lunar craters, with the ages of only 1675 of these determined. This study aims to improve crater identification and age estimation using data from China's Chang'E-1 and Chang'E-2 lunar orbiters. These missions provided high-resolution digital orthophoto images (DOM) and digital elevation models (DEM), offering a rich dataset for analysis. The researchers utilize transfer learning (TL) with deep neural networks (DNNs), a machine learning approach that allows a model trained on one task to be adapted to another, thereby leveraging the limited number of labelled craters (craters with known ages and locations). The process involved a two-stage approach for crater detection and another for age estimation. The work's significance lies in its potential to generate a more comprehensive and accurate lunar crater database, providing better insights into the Moon's geological history and the impact history of the inner Solar System.

Previous research on lunar crater detection employed methods like pattern recognition and machine learning. Deep learning, particularly convolutional neural networks (CNNs), showed promise but was hindered by the limited availability of labelled data representing the full range of crater morphologies. Age estimation relied on methods such as stratigraphic coverage relationships, morphological features (e.g., ray brightness, rim sharpness), optical maturity (OMAT), crater size-frequency distributions (CSFDs), and radiometric dating of returned samples. Each method had limitations; for example, OMAT is best suited for large, rayed craters. This study leverages the strengths of deep learning and transfer learning to overcome the limitations of existing approaches by utilizing a much larger dataset from the Chang'E mission and applying transfer learning techniques for both crater detection and age estimation.

The study employed a two-stage approach for both crater identification and age estimation. For crater identification, the researchers used a multi-scale approach with Chang'E-1 and Chang'E-2 data of varying resolutions (120 m, 50 m, and 7 m). The region-based fully convolutional network (R-FCN) was utilized as the detection model, with ResNet101 serving as the basic network. The first stage involved transfer learning from ImageNet pre-trained parameters and fine-tuning using Chang'E-1 data. The second stage applied this trained model to Chang'E-2 data without further training, representing transductive transfer learning. For age estimation, another two-stage approach was implemented. A dual-channel model was used, one channel processing image data (morphological features from DOM) using a deep CNN and another processing attribute data (morphological and stratigraphic information) using a feedforward neural network. The Mean Teacher semi-supervised learning strategy was incorporated to leverage the large number of newly identified craters. Again, transfer learning was employed, with the model first trained on Chang'E-1 data and then transferred to Chang'E-2 data without retraining. Multiple CNN models were utilized (ResNet50, ResNet101, ResNet152, SENet, etc.), and an ensemble strategy was used to combine the results from these models. The accuracy of the crater detection was evaluated using recall, while the accuracy of age estimation was evaluated using overall accuracy (OA) and confusion matrices.

The two-stage crater detection approach, utilizing transfer learning, achieved high recall rates (94.71% for Chang'E-1 and 93.35% for Chang'E-2), successfully identifying a substantial number of known craters not used in the training set. The approach identified a total of 117,240 craters with diameters ranging from 0.9 to 532 km, a significant increase compared to existing databases. The false-positive rates (FPRs) were low (4.49% ± 0.70% and 4.67% ± 2.10% for craters with diameters of 1-100 km and 100-550 km, respectively). The age estimation model, also using transfer learning, achieved an overall accuracy of 85.44% ± 1.94% (mean ± s.d.) on Chang'E-1 data and correctly classified 89.04% of the dated craters in the Chang'E-2 test set. The analysis of cumulative size-frequency distributions (CSFDs) of craters across different age groups revealed distinct patterns consistent with the established lunar geological time scale. Comparisons with existing databases and studies demonstrated good agreement, although some discrepancies existed, particularly for smaller craters where resolution limitations played a role. The study also provided detailed analysis of crater distributions in specific lunar regions, such as the nearside mare and farside highlands, showing different crater densities and size distributions across these regions, reflecting varying resurfacing events and impact histories.

The findings significantly expand the existing lunar crater database, improving the understanding of the Moon's impact history and geological evolution. The application of transfer learning effectively addressed the challenge of limited labeled data, allowing the researchers to leverage a much larger amount of data for model training. The high recall and low FPR of the crater detection model demonstrate its effectiveness. The consistency of age estimations with existing chronologies supports the accuracy of the age estimation model. Discrepancies observed, particularly for smaller craters, highlight the challenges in detecting and classifying small, degraded craters with the available data resolution. Future research can focus on incorporating higher-resolution data and potentially refining the models to improve the accuracy of small crater detection and age estimation.

This research successfully developed and applied a novel transfer learning approach for the automated identification and age estimation of lunar impact craters. The significantly expanded database and the relatively high accuracy of crater age estimation provide valuable resources for future lunar studies. Limitations regarding the resolution and completeness of the datasets were addressed, and suggestions for future research include using higher-resolution data for better identification of small craters and incorporating more precise stratigraphic data for enhancing age estimation.

The study's limitations stem primarily from the resolution of the Chang'E data used. Smaller craters (< 50 km) were more challenging to detect accurately compared to larger craters. The completeness of the existing dated crater dataset used for age estimation also impacted the accuracy of age assignments. Future work should incorporate higher-resolution data, and more precise stratigraphic data would improve the overall accuracy and completeness of the database.